Abstract
The inherently uncertain and resource-intensive nature of drug discovery is being transformed by machine learning, which accelerates and streamlines the identification of novel therapeutic candidates. This work expands the field of drug discovery and repurposing by employing machine learning techniques to predict compound-protein interactions that may have therapeutic effects, with an emphasis on natural compounds. Due to limited ground truth data for natural product-protein interactions, we used a concordance-based validation approach. We utilized several pre-trained drug-target affinity (DTA) models, including GraphDTA, MLT-LE, and DeepDTA, to predict binding affinities between protein targets listed in the UniProt database and natural products sourced from the COCONUT database. Our analysis demonstrates high concordance indices across all models (significantly exceeding the 0.5 baseline expected for random agreement), indicating strong generalizability to natural compounds despite being trained primarily on small, drug-like compounds. This consistency suggests that these computational approaches can effectively integrate diverse chemical information, thereby enabling a more comprehensive exploration of natural compounds for therapeutic applications.